[mips][FastISel] Implement bswap.

[vkalintiris]

Summary

Implement bswap intrinsic for MIPS FastISel. It's very different for MIPS32 R1/R2.

Based on a patch by Reed Kotler.

[rkotler]

Submit to commits list for comments too.

[dsanders]

test/CodeGen/Mips/Fast-ISel/bswap1.ll

15–16

Please use multiple FileCheck prefixes to handle things that are the same for both commands.

17

As mentioned in D6774, we can significantly simplify this using arguments and returns.

As noted there, I don't mind if this is part of the big cleanup.

51–52

It's possible to shave off one instruction here. At the moment this code is equivalent to this expression:

R = (X << 24) | (X >> 24) | ((X >> 8) & 0xFF00) | ((X << 8) & 0xFF0000)

The '& 0xFF0000' requires an lui+and but with a slight change to the expression we could use an andi instead:

R = (X << 24) | (X >> 24) | ((X >> 8) & 0xFF00) | ((X & 0xFF00) << 8)

67–104

Unnecessary

[vkalintiris]

Rebased, removed redundant instruction & cleaned up the tests.

[dsanders]

LGTM

lib/Target/Mips/MipsFastISel.cpp

bool MipsFastISel::fastLowerCall(CallLoweringInfo &CLI) {
   // Finish off the call including any return values.
   return finishCall(CLI, RetVT, NumBytes);
 }
 
 bool MipsFastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
   switch (II->getIntrinsicID()) {
   default:
     return false;
+  case Intrinsic::bswap: {
+    Type *RetTy = II->getCalledFunction()->getReturnType();
+
+    MVT VT;
+    if (!isTypeSupported(RetTy, VT))
+      return false;
+
+    unsigned SrcReg = getRegForValue(II->getOperand(0));
+    if (SrcReg == 0)
+      return false;
+    unsigned DestReg = createResultReg(&Mips::GPR32RegClass);
+    if (DestReg == 0)
+      return false;
+    if (VT == MVT::i16) {
+      if (Subtarget->hasMips32r2()) {
+        emitInst(Mips::WSBH, DestReg).addReg(SrcReg);
+        updateValueMap(II, DestReg);
+        return true;
+      } else {
+        unsigned TempReg[3];
+        for (int i = 0; i < 3; i++) {
+          TempReg[i] = createResultReg(&Mips::GPR32RegClass);
+          if (TempReg[i] == 0)
+            return false;
+        }
+        emitInst(Mips::SLL, TempReg[0]).addReg(SrcReg).addImm(8);
+        emitInst(Mips::SRL, TempReg[1]).addReg(SrcReg).addImm(8);
+        emitInst(Mips::OR, TempReg[2]).addReg(TempReg[0]).addReg(TempReg[1]);
+        emitInst(Mips::ANDi, DestReg).addReg(TempReg[2]).addImm(0xFFFF);
+        updateValueMap(II, DestReg);
+        return true;
+      }
+    } else if (VT == MVT::i32) {
+      if (Subtarget->hasMips32r2()) {
+        unsigned TempReg = createResultReg(&Mips::GPR32RegClass);
+        emitInst(Mips::WSBH, TempReg).addReg(SrcReg);
+        emitInst(Mips::ROTR, DestReg).addReg(TempReg).addImm(16);
+        updateValueMap(II, DestReg);
+        return true;
+      } else {
+        unsigned TempReg[8];
+        for (int i = 0; i < 8; i++) {
+          TempReg[i] = createResultReg(&Mips::GPR32RegClass);
+          if (TempReg[i] == 0)
+            return false;
+        }
+
+        emitInst(Mips::SRL, TempReg[0]).addReg(SrcReg).addImm(8);
+        emitInst(Mips::SRL, TempReg[1]).addReg(SrcReg).addImm(24);
+        emitInst(Mips::ANDi, TempReg[2]).addReg(TempReg[0]).addImm(0xFF00);
+        emitInst(Mips::OR, TempReg[3]).addReg(TempReg[1]).addReg(TempReg[2]);
+
+        emitInst(Mips::ANDi, TempReg[4]).addReg(SrcReg).addImm(0xFF00);
+        emitInst(Mips::SLL, TempReg[5]).addReg(TempReg[4]).addImm(8);
+
+        emitInst(Mips::SLL, TempReg[6]).addReg(SrcReg).addImm(24);
+        emitInst(Mips::OR, TempReg[7]).addReg(TempReg[3]).addReg(TempReg[5]);
+        emitInst(Mips::OR, DestReg).addReg(TempReg[6]).addReg(TempReg[7]);
+        updateValueMap(II, DestReg);
+        return true;
+      }
+    }
+    return false;
+  }
   case Intrinsic::memcpy:
   case Intrinsic::memmove: {
     const auto *MTI = cast<MemTransferInst>(II);
     // Don't handle volatile.
     if (MTI->isVolatile())
       return false;
     if (!MTI->getLength()->getType()->isIntegerTy(32))
       return false;

test/CodeGen/Mips/Fast-ISel/bswap1.ll

+; RUN: llc < %s -march=mipsel -mcpu=mips32 -O0 -relocation-model=pic \
+; RUN:      -fast-isel=true -mips-fast-isel -fast-isel-abort=1 | FileCheck %s \
+; RUN:      -check-prefix=ALL -check-prefix=32R1
+; RUN: llc < %s -march=mipsel -mcpu=mips32r2 -O0 -relocation-model=pic \
+; RUN:      -fast-isel=true -mips-fast-isel -fast-isel-abort=1 | FileCheck %s \
+; RUN:      -check-prefix=ALL -check-prefix=32R2
+
+@a = global i16 -21829, align 2
+@b = global i32 -1430532899, align 4
+@a1 = common global i16 0, align 2
+@b1 = common global i32 0, align 4
+
+declare i16 @llvm.bswap.i16(i16)
+declare i32 @llvm.bswap.i32(i32)
+
+define void @b16() {
+  ; ALL-LABEL:  b16:
+
+  ; ALL:            lw    $[[A_ADDR:[0-9]+]], %got(a)($[[GOT_ADDR:[0-9]+]])
+  ; ALL:            lhu   $[[A_VAL:[0-9]+]], 0($[[A_ADDR]])
+
+  ; 32R1:           sll   $[[TMP1:[0-9]+]], $[[A_VAL]], 8
+  ; 32R1:           srl   $[[TMP2:[0-9]+]], $[[A_VAL]], 8
+  ; 32R1:           or    $[[TMP3:[0-9]+]], $[[TMP1]], $[[TMP2]]
+  ; 32R1:           andi  $[[TMP4:[0-9]+]], $[[TMP3]], 65535
+
+  ; 32R2:           wsbh  $[[RESULT:[0-9]+]], $[[A_VAL]]
+
+  %1 = load i16, i16* @a, align 2
+  %2 = call i16 @llvm.bswap.i16(i16 %1)
+  store i16 %2, i16* @a1, align 2
+  ret void
+}
+
+define void @b32() {
+  ; ALL-LABEL:  b32:
+
+  ; ALL:            lw    $[[B_ADDR:[0-9]+]], %got(b)($[[GOT_ADDR:[0-9]+]])
+  ; ALL:            lw    $[[B_VAL:[0-9]+]], 0($[[B_ADDR]])
+
+  ; 32R1:           srl   $[[TMP1:[0-9]+]], $[[B_VAL]], 8
+  ; 32R1:           srl   $[[TMP2:[0-9]+]], $[[B_VAL]], 24
+  ; 32R1:           andi  $[[TMP3:[0-9]+]], $[[TMP1]], 65280
+  ; 32R1:           or    $[[TMP4:[0-9]+]], $[[TMP2]], $[[TMP3]]
+  ; 32R1:           andi  $[[TMP5:[0-9]+]], $[[B_VAL]], 65280
+  ; 32R1:           sll   $[[TMP6:[0-9]+]], $[[TMP5]], 8
+  ; 32R1:           sll   $[[TMP7:[0-9]+]], $[[B_VAL]], 24
+  ; 32R1:           or    $[[TMP8:[0-9]+]], $[[TMP4]], $[[TMP6]]
+  ; 32R1:           or    $[[RESULT:[0-9]+]], $[[TMP7]], $[[TMP8]]
+
+  ; 32R2:           wsbh  $[[TMP:[0-9]+]], $[[B_VAL]]
+  ; 32R2:           rotr  $[[RESULT:[0-9]+]], $[[TMP]], 16
+
+  %1 = load i32, i32* @b, align 4
+  %2 = call i32 @llvm.bswap.i32(i32 %1)
+  store i32 %2, i32* @b1, align 4
+  ret void
+}